May 19 (Tues) @ 3:00pm: "Physics-inspired Machine Learning with Probabilistic Computers," Shaila Niazi, ECE PhD Defense

Location: Zoom Link — https://ucsb.zoom.us/j/88333698948

Abstract

Energy-based models offer immense representational capacity for both generative artificial intelligence and complex quantum many-body physics, yet scaling them on conventional, deterministic hardware remains a fundamental computational challenge. This dissertation advances probabilistic computing through hardware-algorithm co-design, utilizing field-programmable gate arrays (FPGAs) to physically emulate massively parallel networks of stochastic bits (p-bits).

First, the dissertation demonstrates that hardware-aware sparse deep Boltzmann machines (DBMs) can be efficiently trained for classical machine learning tasks. By mapping sparse DBMs to FPGA, the system matches the classification accuracy of dense restricted Boltzmann machines using two orders of magnitude fewer parameters. Furthermore, this deep probabilistic framework unlocks generative synthesis capabilities, effortlessly generating novel images where highly parameterized shallow architectures fail.

Building upon this foundation, the framework is extended to quantum simulation by utilizing the classical Boltzmann distribution to natively represent Neural Quantum States (NQS). A novel dual-sampling algorithm is introduced that replaces intractable marginalization with conditional sampling, enabling the stable training of deep NQS. Implemented natively on an FPGA, this approach successfully trains deep networks for the 2D transverse-field Ising model at chemical accuracy, providing a systematically improvable path for scalable quantum simulation.

Bio

Shaila Niazi is a Ph.D. candidate in the Electrical and Computer Engineering Department at UC Santa Barbara, advised by Prof. Kerem Y. Camsari. She earned her B.S. in Electrical and Electronic Engineering from the Bangladesh University of Engineering and Technology (BUET) in 2017 and her M.S. in ECE from UCSB in 2024. Her research focuses on hardware–algorithm co-design for probabilistic computing, with applications in accelerating deep energy-based model training for both generative AI and quantum simulations.

Hosted By: Professor Kerem Y. Camsari

Submitted By: Shaila Niazi <sniazi@ucsb.edu>